Research map / Video / Video-language understanding
FrameThinker: Learning to Think with Long Videos via Multi-Turn Frame Spotlighting
Zefeng He, Xiaoye Qu, Yafu Li, Siyuan Huang, Daizong Liu, Yu Cheng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace uniform 32-frame sampling with multi-turn frame spotlighting | Qwen2.5-VL-7B · LongVideo-Reason | accuracy (%): 64.1 -> 76.1 | Improved |
| train with FrameThinker (SFT+RL) on Video-Holmes | Qwen2.5-VL-7B · Video-Holmes | accuracy (%): 27.8 -> 56.1 | Improved |
| large-scale FrameThinker training | Qwen2.5-VL-7B · six benchmarks average | accuracy (%): +10.4 | Improved |
| FrameThinker versus Video-R1 on MLVU | Video-R1 · MLVU | accuracy (%): 60.2 -> 59.1 | Hurt |
| add a format reward (ablation) | FrameThinker RL · Video-Holmes | action usage | Hurt |
| use unconditional action bonus instead of conditional (ablation) | FrameThinker RL · Video-Holmes | training stability | Hurt |
| add Cognitive Consistency Verification (CCV) in training and inference (ablation) | FrameThinker · Video-Holmes | accuracy | Improved |
| larger bonus for get frame number than choose frames (0.2 vs 0) (ablation) | FrameThinker RL · Video-Holmes | accuracy | Improved |
| reward more turns, R_action = k(T-1) (ablation) | FrameThinker RL · Video-Holmes | training stability | Hurt |
| apply FrameThinker SFT/RL setup to baseline Qwen2.5-VL-7B versus full approach (ablation) | FrameThinker · Video-Holmes | accuracy | Improved |
Later papers that compared with this method
2 later accepted papers report FrameThinker as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- VideoBrain: Learning Adaptive Frame Sampling for Long Video Understanding (ICML 2026 · better on 3 of 3 shared results)
- EVA: Efficient Reinforcement Learning for End-to-End Video Agent (CVPR 2026 · better on 3 of 3 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- When Thinking Drifts: Evidential Grounding for Robust Video Reasoning (NeurIPS 2025 · similarity 0.91)
- Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.91)
- VideoReasonBench: Can MLLMs Perform Vision-Centric Complex Video Reasoning? (ICLR 2026 · similarity 0.91)
- OneThinker: All-in-one Reasoning Model for Image and Video (CVPR 2026 · similarity 0.91)
- Scaling RL to Long Videos (NeurIPS 2025 · similarity 0.91)
- Think-as-You-See: Streaming Chain-of-Thought Reasoning for Large Vision-Language Models (CVPR 2026 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
